Unverified50% confidenceOpinionExact time
On contaminated benchmarks, models can directly extract answers from parametric memory rather than relying on in-context retrieval, rendering the reported lossless-compression conclusions meaningless.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
Unverified提示缓存缓存的是模型处理输入时产生的中间计算结果（KV对），而非模型的输出73% similarUnverifiedA model's commitments and memory writes do not change its underlying probability distribution, and its reflection is essentially just generating text that fits conversational expectations rather than truly updating behavioral strategies, so it may repeat the same mistakes.71% similarUnverified该系统在提示词中明确要求模型只能使用提供的上下文作答，答案不在上下文中时直接回答不知道，而非凭训练记忆编造71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/828097API
curl https://kongchang.com/api/v1/knowledge/claims/828097MCP
get_claim(id=828097)